Predicting Clinical Outcomes with Patient Stratification via Deep Mixture Neural Networks

The increasing availability of electronic health record data offers unprecedented opportunities for predictive modeling in healthcare informatics including outcomes such as mortality and disease diagnosis as well as risk factor identification. Recently, deep neural networks (DNNs) have been successfully applied in healthcare informatics and achieved state-of-art predictive performance. However, existing DNN models either rely on the pre-defined patient subgroups or take the "one-size-fits-all" approach and are built without considering patient stratification. Consequently, those models are not able to discover patient subgroups and the risk factors are thereafter identified for the entire patient population, failing to account for potential group differences. To address this challenge, we propose the use of deep mixture neural networks (DMNN), a unified DNN model for simultaneous patient stratification and predictive modeling. Experimental results on a clinic dataset show that our proposed DMNN can achieve good performance on predicting diagnosis of acute heart failure. With DMNN's ability to incorporate patient stratification, we are able to systematically identify group-specific risk factors for different patient subgroups which could potentially shed light on revealing factors that contribute to outcome differences.